With the rapid development of big data and artificial intelligence， deep learning-based natural language processing has achieved remarkable advances in text processing. Educational policy documents are important research materials with great analytical potential. Nevertheless， restricted by standardized formats， unified contents and inherent long-tail distribution characteristics， current neural network-based topic mining approaches still have room for improvement. This study presents a neural topic model based on probability cluster regularization （PCR-NTM）， which aims to improve the performance of topic modeling for educational policies and alleviate the topic folding issue. Experimental results demonstrate that the proposed method outperforms existing state-of-the-art neural topic models in terms of topic coherence， topic diversity and normalized mutual information. It also achieves substantial gains in precision， recall and F1-score. Furthermore， this paper explores the development trends and research hotspots of educational policies in recent years， which provides a useful reference for policy analysis.